Combining point and distributed strain sensor for complementary data-fusion: A multi-fidelity approach
Seung-Seop Jin, Sung Tae Kim, Young-Hwan Park
Abstract
The aim of this paper is to develop a complementary data-fusion algorithm for monitoring of the strain over whole structural system. Point strain sensors measure accurate strains (high-accuracy) at discrete measurement positions (low-spatial resolution), while a distributed strain sensor enables a quasi-continuous distributed measurement (high-spatial resolution) with less accurate strains (low-accuracy). This study firstly investigates the complementary data-fusion for the point and distributed strain sensor to combine their advantages in order to obtain the accurate strain distribution (i.e., high-accuracy and high-spatial resolution). For this purpose, a traditional multi-fidelity data-fusion framework can be applied based on Gaussian process regression in the form of auto-regressive scheme. However, the traditional method is not efficient to learn complex cross-correlations. To enable flexible learning for the complex cross-correlations, this study introduces a multi-fidelity data-fusion framework using the input-connected Gaussian process mapping. The numerical and experimental studies demonstrate that the proposed method outperforms the traditional methods with more accurate predictions by using fewer samples for the high-fidelity data. In this context, the proposed method has potential for monitoring the strain distribution over whole structural systems under limited budgets.